Throughout 2022, artificial intelligence was in the spotlight. From DALL-E to ChatGPT to MidJourney, generative AI stunned the world. Here’s a look back at a year marked by technology, which could be just a foretaste of the revolution to come in 2023…
From the war in Ukraine to the historic World Cup final, 2022 was a busy year. However, it was also a year marked by the rise of a technology: generative artificial intelligence.
Over the past twelve months, we’ve watched in amazement as AI has expanded rapidly across the development of huge language models all over the world. A veritable revolution has begun, and 2022 may well be remembered as the dawn of a new era…
The most significant AIs of 2022
Thousands of artificial intelligence models were deployed in 2022. However, only a handful of them made headlines. Here’s a selection of the most impressive AIs and useful AIs launched this year…
Relaxed in November 2022, OpenAI text-davinci-003 is an open language model. It is the new default engine for GPT-3. Its most surprising feature is its ability to rhyme.
Another striking language model is Google PaLM5, which completes the Pathways family.. Unlike text-davinci-003, however, this is a closed model.
Of course, we can’t go back to 2022 without mentioning MidJourney: the generative AI model ” text-to-image “capable of generating artistic images from text. The V4 version deployed in November 2022 is particularly impressive.
Other Text-to-Image models released during the year include OpenAI’s DALL-E and Stable Diffusion. For its part, Google launched Parti trained on a dataset of a record 4.8 billion images. However, this is a closed model reserved for researchers.
On November 30, 2022, OpenAI signed the show-stopper with ChatGPT: a chatbot based on GPT-3, freely available on the web, capable of answering any question in natural language…
The new AI tycoons
The year 2022 was also marked by the success of AI companies. Several companies achieved a capitalization of several billion dollars thanks to their AI models. However, many investments remain confidential…
Industry winners include Sonantic, having raised $95 millionand Xiaoice, which reached a capitalization of $1,000 million thanks to $138 million in investment.
For its part, Stability.ai raises $100 million and is now valued at $1,100 million. It owes its success in particular to the open source text-to-image artificial intelligence model Stable Diffusion.
Jasper.ai, an artificial intelligence for SEO copywriting and marketing, has raised $125 million. It is now valued at $1,500 million. Find out whether Jasper.ai can replace a human copywriter in our full test at this address. On its side, Hugging Face raises $100 million in investment and reached a valuation of $2,000 million.
The the big winner of 2022 is, of course, OpenAI. The Californian company founded in 2015 by Elon Musk and Sam Altman has reached a capitalization of $20,000 million, thanks in particular to the hype generated by DALL-E and ChatGPT…
Alphabet’s (Google) DeepMind artificial intelligence branch, meanwhile, has raised $1,200 million. In this case, it’s actually the costs borne by Alphabet. Its new models have been adopted by researchers, in particular to offer personalized experiences.
For example, the Chinchilla model has been used to create Dramatron capable of creating theater scripts. At the Fringe Festival in August 2022, human actors performed the plays generated by this AI.
Applications based on these models cover a wide range of domains, and include many startups use them to create innovative platforms and tools. However, only a handful of players are creating new models.
The year of Text-to-Image
One category of artificial intelligence has seen a particularly meteoric rise in 2022: text-to-image models. From a few key words entered by the user, these AIs can generate totally new images.
Monetization of these models has been much faster than expected. For example, toy giant Mattel has used OpenAI’s DALL-E 2 to design new cars for its Hot Weels range. In particular, it uses AI to test different colors or alter a design. According to the company, the aim is to amplify the quality of ideas and come up with new ones.
Similarly, in November 2022, German brewer Brauquadrat used Midjourney V4 to generate artistic images for its sour beer range. Starting with a simple prompt “commercial raspberry photo, teal background, splash, juicy”, the AI generated striking results.
At the same time, Stability.ai announced the adoption of its Stable Diffusion model by more than 200,000 software developers. A major milestone for the largest open-source Text-to-Image model.
In addition to DALL-E 2 and Stable Diffusion, numerous Text-to-Image models were released in 2022. These AIs are now integrated into the most popular softwaresuch as Canva, DeviantArt, AutoCAD, Photoshop and Lensa’s new Magic Avatars tool.
Increasingly large models
Text-to-Image models are becoming broader and broader, and the most recent ones are trained on billions of image and text pairs for the equivalent of hundreds of years. There are several categories of models: autoregressive, diffusion and GAN (generative adversarial networks). Here are a few examples of models in each category, and their number of parameters.
- OpenAI DALL-E 1, January 2021
- Tsinghua CogView, May 2021, 4 billion parameters
- Google Parti, June 2022, 20 billion parameters
- Tsinghua CogView 2, June 2022, 24 billion parameters
- Microsoft NUWA-Inifinity, July 2022
- Midjourney v1, April 2022
- OpenAI DALL-E 2, April 2022, 1 million users in 3 months
- Google Imagen, May 2022
- Stability.ai Stable Diffusion, August 2022, 1 million users in 50 countries
- Baidu ERNIE-ViLG 2.0, October 2022, 24 billion parameters
- Nvidia eDiff-I, November 2022
- MidJourney V4, November 2022
- CrAIyon (DALL-E Mini), August 2022, 2.5 billion parameters
Let’s also mention Text-to-Video” models such as Google Imagen Video and Phenaki, or Meta Make-A-Video. As their name suggests, these AIs go a step further by creating videos from text entered by the user. In the future, they could generate your favorite future movie or series in an instant, or create virtual reality worlds…
Language models such as GPT are also becoming increasingly broad, and could be combined with Text-to-Image to create a huge multimodal model. If we look at the ranking of the largest language models, we can see that China and Russia will soon rival the United States in the open-source field…
The rise of open-source
More 1000 researchers from 60 countries replicated GPT-3 using more than 40 languages. Training of the model began on March 11, 2022 and was completed on July 6, 2022 using 384 A100 GPUs on the French public supercomputer Jean Zay at a total cost of $7 million. Initially entitled tr11-176B-ml, this open-source model has been eventually named BLOOM.
In parallel, Google’s Russian equivalent, Yandexreleased a 100-billion-parameter model capable of speaking Russian and English. China, meanwhile, has opened up its GLM-130B model to the rest of the world, a move taken up by WeChat with WeLM 10B.
In the USA, Amazon has promised to open its new Alexa Teacher Model (AlexaTM 20B) and Meta also opened several models to the public. For example, a Galactica demo (GAL 120B) was presented to the public. Unfortunately, abuse and hijacking by malicious actors prompted Meta to withdraw this demo.
When AI enters the2s business world
From many Fortune 500 companies use AI models. This applies to all business sectors, including healthcare, retail, engineering and fashion.
If we using GPT-3 as an exampleThis model, created by OpenAI, is used by Microsoft, Shell, Morgani Stanley, IBM, HSBC, PWC, EY, Accenture, AON, Cognizant, WiPro, Cisco, Intel, Salesforce, Disney, BMW, Jasper.AI, Autodesk and Twitter, among others.
The explosion of hardware and data
In order to design their new chipsmanufacturers are now using AI. This is the case with Google’s TPUs, and Nvidia’s Hopper H100 GPUs. But AI-designed circuits are smaller and faster.
This has made it possible to considerably increase chip powerin turn used to train AI models. For example, it is common to combine several thousand TPUs to train a model in parallel during the equivalent of hundreds of years.
Thus, the new GPUs Hopper H100 are six times faster than the A100 chips used to drive most AI models in 2022. We can therefore expect even more impressive models as early as 2023…
Furthermore, in March 2022, Jordan Hoffmann and 21 other DeepMind researchers shared a major discovery. They realized that AI labs had only been using 9% of the of the volume of data they should have used to train their models.
Thus, while GPT-3 175B was trained on around 300 billion tokens (about 600 gigabits), it should have been trained on 3,500 billion tokens or the equivalent of 7 terabits. As a result, laboratories began to collect more and more data.
Some have even exceeded the recommended data volumeWeChat WeLM 10B, Amazon AlexaTM 20B and Microsoft Z-Code++. Others have adopted an original approach, such as OpenAI, whose transforming speech recognition model Whisper has been trained on 77 years of audio content. This AI could help create new text datasets based on audio, including speech extracted from video.
AI on the verge of surpassing humans?
The latest AI models are outperforming humans on school exams. For example, Google Pathways models have obtained a higher score than most students on mathematics tests.
The Minerva 540B model from July 2022 scored 14% above average on Poland’s national mathematics exam, and 79% above average on the UK national mathematics exam..
Similarly, PaLM 540B dated April 2022 outperforms humans on SuperGLUE benchmark and Flan-PaLM from October 2022 performs twice as well as the human average on the MMLU benchmark.
This rapid evolution makes it possible to create numerous applications for language models. The creator of Google Transformer, Dr Ashish Vaswani, founded Adept to apply this technology to any web browser or computer task.
With a budget of $65 million and the help of researchers from DeepMind, Google Brain and OpenAI, the team has released its first Action Transformer model (ACT-1) in September 2022. This model is currently attached to a Chrome extension to enable it to observe what’s happening on the browser and perform certain actions such as clicking, typing or scrolling.
ACT-1 can be used to monitor real estate search instructionsadd customers to a CRM, write e-mails, and much more. The results are impressive, and future iterations will be able to interact with various applications on a PC.
The future of AI in 2023
Several major new developments are expected in AI in 2023. First of all, DeepMind is already training its new generalist agent Gato. It could be the first truly general artificial intelligence.
The Google Pathways family to expandto cover the 1000 most widely spoken languages in the world. This will enable greater inclusion of billions of people from minority backgrounds around the world.
In addition, three years after GPT-3, OpenAI could release GPT-4. This new language model is likely to shake up the current AI landscape, with a new technological leap forward…
And if 2022 was the year of Text-to-Image models, 2023 could be the year of Text-to-Video models. Several models of this type were unveiled during the year, but they should now reach a milestone in terms of definition and refresh rate.
In the future, general artificial intelligence could solve several major problems for mankind. Autonomous vehicles could put an end to manual driving and the fatal accidents that occur every 24 seconds worldwide.
General AI could also free us from work, since it will be 99% more productive than humans. It would also be possible to generate customized diets to eliminate obesity and malnutrition.
However, it will undoubtedly be necessary many years to come for the tremendous potential of AI to be applied to the eight billion people on this planet. And while major companies are already exploiting this technology, governments and public institutions are lagging far behind.
Be that as it may, the revolution has begun and will continue in 2023 with the emergence of spectacular new models. The road to general-purpose AI is now mapped out, and there’s no stopping the advent of this new era.…